Tuesday, Oct 13

Brad Smith (Microsoft) said this on Oct 12, 2026; not an official company announcement (source: techcrunch.com).

On a BBC podcast, Microsoft President Brad Smith called for urgent AI regulation, arguing that milk and washing machines face more oversight than AI companies. He said the UK, whose AI Security Institute gets pre-deployment access to frontier models, is better placed than the US or EU to lead global safety rules. Smith called the White House's voluntary safety pledge a good step but not enough, and backed breach-notification duties for rogue AI agents, outside evaluator approval before release, and kill switches.

Not confirmed by the company — reported on Oct 12, 2026 (source: techcrunch.com).

Santa Monica-based Blueprints AI, founded by architect Al Drams and data scientist Coco Cao, uses AI to produce code-compliant, permit-ready architectural blueprints from a 2D sketch or PDF. The company, a 2026 Startup Battlefield 200 pick, plans to unveil a second-generation version of its software at TechCrunch Disrupt, held October 13 to 15 in San Francisco. Drams says architects spend about 80% of their time on drafting and wants to flip that toward design.

Posted on arXiv on Oct 12, 2026 (source: arxiv.org).

A Microsoft Research study finds long-running autoresearch agents get stuck in 'idea basins' because accumulated chat context keeps steering them toward similar solutions, even with more compute. Fork-and-Flush splits an agent into parallel branches that keep the workspace and files but start with an empty chat context. Across 13 long-running tasks using GPT-5.6-terra and GPT-5.5 via the GitHub Copilot CLI, it scored 0.78 normalized versus 0.47 for single runs and 0.54 for best-of-4 on the same compute, per the paper.

Posted on arXiv on Oct 12, 2026 (source: arxiv.org).

Google DeepMind's AlphaProof Nexus uses LLM agents to search for formal proofs in Lean, with the Lean compiler checking every step. The peer-reviewed paper, published in Science on October 8 after a May arXiv preprint, reports autonomously solving 9 of 353 attempted Erdos problems, including two open for 56 years, and proving 44 of 492 open OEIS conjectures. DeepMind says the full agent cost a few hundred dollars of inference per problem and also helped on problems in algebraic geometry, optimization, quantum optics and graph theory.

Posted on arXiv on Oct 12, 2026 (source: arxiv.org).

U-Space maps semantic anchors for doubt and certainty into a model's residual space, covering ambiguity, incomplete information, conflicting evidence and general uncertainty. Its U-Lens score combines how strongly each token's hidden state leans toward doubt with predictive entropy, from the one generation that produced the answer, with no correctness labels, repeated sampling or training. The authors say it beats established baselines on reasoning benchmarks, including length-controlled evaluation; code is on GitHub.

Hugging Face published this on Oct 13, 2026 (source: huggingface.co).

Fine-grained access tokens and service account tokens can now get a contributor permission that lets them create repositories in your account or selected organizations, with full write access only to the repositories they created. Hugging Face says this suits CI/CD pipelines, training jobs and agents that publish new checkpoints or datasets without write access to the whole namespace.

Anthropic published this on Oct 13, 2026 (source: anthropic.com).

Anthropic researchers ran Gemma4-26B-A4B split across 15 off-brand Android TV boxes (about $55 each) at 0.39 tokens per second to show that a botnet of weak devices could host a capable model. In their evals, a 9B model got inference running on 8 of 9 hardware types, and GLM-5.2 and stronger models completed end-to-end replication in simplified tests. Models of 30B parameters and under can already do many individual replication steps but make mistakes chaining them. Anthropic estimates small models could reach this level by the end of the year if they keep improving at their historical rate, and closes with suggested defenses.

Not confirmed by the company — reported on Oct 12, 2026 (source: techcrunch.com).

The US Air Force will buy atmospheric data from WindBorne Systems' fleet of about 700 autonomous weather balloons, which also feeds the startup's own deep-learning forecasting model, WeatherMesh 6. The contract is worth up to $49 million, is paid per order, and was awarded without competitive bidding; $1.7 million has been ordered so far. CTO Andrey Sushko said it is the largest contract the Redwood City company has been part of.

Not confirmed by the company — reported on Oct 12, 2026 (source: techcrunch.com).

Honeyjar, founded by PR veterans Michelle Masek and Nadia Jamshidi, automates communications work such as writing announcements, finding journalists to pitch, tracking coverage and spotting speaking opportunities. The founders say the goal is to bring the whole comms workflow together for founders, marketers and in-house teams rather than replace PR professionals. Honeyjar says its waitlist drew 10,000 signups in the first 24 hours, with beta testers including Nextdoor, Coco Robotics and Bloomberg Beta. The startup is in TechCrunch Disrupt's Startup Battlefield 200.

AWS published this on Oct 12, 2026 (source: aws.amazon.com).

AWS says GPT-6 Sol and GPT-6 Luna are available through global cross-Region inference from all five AWS Regions, and GPT-6 Astra from Asia Pacific (Singapore), so teams there can go beyond single-Region capacity when demand would otherwise throttle. The post shows calls through the Converse API, the OpenAI-compatible APIs and coding clients on the bedrock-runtime endpoint. It also covers IAM and SCP permissions, quota sizing and monitoring inference from the source Region.

Hugging Face (OpenBMB) published this on Oct 12, 2026 (source: huggingface.co).

OpenBMB released model cards for MiniCPM-V 4.7 under Apache 2.0, in a 1B variant and a 35B-A3B mixture-of-experts variant for single-image, multi-image and video understanding. It compresses visual tokens early inside the vision encoder and supports mixed 4x/16x visual token compression. OpenBMB says the 1B model beats Qwen3.5 0.8B in Instruct mode on overall score and inference efficiency. The models run on iOS, Android and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD and Arm chips and for SGLang, vLLM, llama.cpp and Ollama.

Micah Lee said this on Oct 12, 2026; not an official company announcement (source: micahflee.com).

In a blog post, Micah Lee argues that many open-weight models are about as good as the latest ChatGPT and Claude models, including for vulnerability research. He says they can run privately on your own hardware or through enclave-based providers like Tinfoil, and that switching providers is easy. Lee argues ChatGPT and Claude will never be private and criticizes their cyber guardrails and vetted access programs as corporate gatekeeping.

Not confirmed by the company — reported on Oct 12, 2026 (source: arstechnica.com).

Harvard researchers Fiona Chen and James Stratton analyzed Jellyfish data covering 300 million work events from more than 700,000 employees at over 700 software firms from 2021 through March 2026. After firms adopted AI coding agents, lines of code rose 30%, commits 20% and pull requests 23% on average. Resolution rates for Jira issues and epics did not change significantly, because code review took longer, pull requests needed more revisions and reviewers left more comments. The authors found little evidence that firms increased software output or reduced employment.

Ollama published this on Oct 13, 2026 (source: github.com).

The v0.40.4-rc0 pre-release updates the bundled llama.cpp to b11521 for decision models and EmbeddingGemma 2. The server now reuses verified manifests and loaded models in inference routes, and the MLX backend compacts restored recurrent state after speculative rollback. Ollama's launch command also includes installed models in the Codex /model picker.

Cognition published this on Oct 13, 2026 (source: docs.devin.ai).

Cognition's Devin release notes say Devin now remembers what it learns about each user and their work across sessions. Memory is personal and not shared with the organization; users can review its files and turn it on or off in Settings, and each night Devin reviews recent sessions to refine what it remembers. Org admins can turn memory on or off, and enterprise admins can force it for every organization.

Runway published this on Oct 13, 2026 (source: docs.dev.runwayml.com).

Runway's API changelog says Ideogram 4.5 (ideogram_4_5) is now available on Runway Dev, generating images with legible typography and layout from prompts of up to 10,000 characters, with up to five reference images and one to eight outputs. It supports precise edits and mask-based edits. Quality tiers cost 1, 4, 8 and 25 credits per image.

@cognition on X published this on Oct 12, 2026 (source: x.com).

Artificial Analysis launched a leaderboard for FrontierCode 1.1, Cognition's private benchmark that tests whether a model's changes to real open source repositories would be merged by maintainers. Artificial Analysis says Claude Opus 5.5 leads at 52%, nine models from six labs score between 43% and 52%, and MiMo-V2.6-Pro is the top open-weights model at 45%. It says GPT-6.1 Sol scores 49% at $0.81 per task, matching Claude Fable 5.1 at under a tenth of the cost.

@googlecloud on X published this on Oct 12, 2026 (source: x.com).

Google Cloud shared a Gemini at Work data demo by Gabe Weiss that goes from Lakehouse discovery to a trained XGBoost model and a live BI dashboard in four minutes. The demo uses live PySpark with zero-copy joins on Lightning Engine, explainable XGBoost on Managed Spark at 90% accuracy, and BI refreshes via Knowledge Catalog.

@OpenAIDevs on X published this on Oct 12, 2026 (source: x.com).

OpenAI Developers shared a demo by @kagigz showing custom annotation controls and suggested prompts paired with WebMCP. OpenAI says the combination gives agents more context to act on, letting users edit scenes and update expenses in ChatGPT's browser.

@NVIDIAAI on X published this on Oct 12, 2026 (source: x.com).

NVIDIA AI said the Fin team fine-tuned Nemotron 3 Super and Nano for its new Fin Apex customer service models. Fin's Fergal Reid announced the Fin Apex 2.0 series, saying the models are fast and efficient and that in a backtest they outperformed the general model Astra on Fin's tasks.

@nvidia on X published this on Oct 12, 2026 (source: x.com).

NVIDIA announced launch events for its RTX Spark Windows laptops in Beijing, Berlin, Istanbul, Seoul, San Francisco and Taipei, with demos, swag and a chance to win an ASUS ProArt P14 RTX Spark laptop signed by Jensen Huang. NVIDIA says event details are coming soon.

@MiniMax_AI on X published this on Oct 12, 2026 (source: x.com).

MiniMax said Linda Sheng, its President of Global Business, spoke at the inaugural DBS Asia Future Forward Conference 2026 in Singapore. Other speakers included Singapore Prime Minister Lawrence Wong and TikTok CEO Shou Chew.

@perplexity_ai on X published this on Oct 12, 2026 (source: x.com).

Perplexity published a post on how it built the filesystem behind Perplexity Projects as a Git forge for agents. Each agent works in its own sandbox and commits to a shared repository, so people and other agents can build on its work across sessions.

@OpenAIDevs on X published this on Oct 12, 2026 (source: x.com).

OpenAI Developers said applications to become a Codex Ambassador are open again. The program supports community members who host meetups or share what they have learned building with Codex.

Alexandr Wang (Meta) said this on Oct 12, 2026; not an official company announcement (source: x.com).

Alexandr Wang quoted an early Muse Wonder tester who said that on every technical benchmark they ran, the model was neck and neck with or better than GPT-6 Astra, Claude Fable and Claude Opus. The tester also wrote that Meta is going to win consumer agents in a landslide.

@GoogleWorkspace on X published this on Oct 12, 2026 (source: x.com).

Google Workspace promoted how small businesses in the US use Gemini in Workspace to multiply daily output, linking to its guide on getting started with Google AI for work.

@googlecloud on X published this on Oct 12, 2026 (source: x.com).

Google Cloud said Honeywell is partnering with it to bring AI into industrial operations. Gemini connects to control systems, analyzes operational health in real time and predicts failures before they cause downtime.

@nvidianewsroom on X published this on Oct 12, 2026 (source: x.com).

NVIDIA and Cisco introduced USFP pluggable optics, which NVIDIA says is the first module to support 400Gb/s per lane, with four times the bandwidth density of current optics. NVIDIA says builders can connect more compute to the same switch without extra network tiers and with less power per bit.

@TencentGlobal on X published this on Oct 12, 2026 (source: x.com).

Tencent promoted WorkBuddy, saying it brings AI into the files and tools where work already happens, using the example of a quarterly review that spans six apps.

@GoogleWorkspace on X published this on Oct 12, 2026 (source: x.com).

Google Workspace said teams can ask Gemini to turn raw ideas into a collaborative workshop board inside a Sheets canvas, where teammates post categorized cards, upvote ideas and assign action items in real time. Sheets canvas turns spreadsheet data into interactive views such as Kanban boards, dashboards and calendars.

@Kling_ai on X published this on Oct 12, 2026 (source: x.com).

Kling AI released a Creator Off Script episode with Simon Meyer, creative director and co-founder of PROMPTR Studios, who got early access to Kling 4.0 and made a music video that went viral. He walks through his process, how he makes AI footage feel real, and the Kling 4.0 features that impressed him most.

@UnrealEngine on X published this on Oct 12, 2026 (source: x.com).

Epic's Unreal Engine said its EMEA Dev Days run this month in Poland, Germany, Sweden and France, with a technical program covering Unreal Engine 5.8 features, emerging AI, MCP workflows and MetaHuman characters.

@alibaba_cloud on X published this on Oct 12, 2026 (source: x.com).

At Apsara Conference 2026, Qoder senior staff engineer William Yao said Qoder tackles coding bottlenecks by building a knowledge layer that connects codebase information, memory and task context across development workflows.

@alibaba_cloud on X published this on Oct 12, 2026 (source: x.com).

At Apsara Conference 2026, Alibaba Cloud's head of cloud-native application platform Qi Zhou discussed putting native AI agents to work in enterprises. He said AgentCore helps teams coordinate and evaluate agents and turns validated improvements into reusable capabilities for future tasks.

@Huawei on X published this on Oct 12, 2026 (source: x.com).

Huawei and its customers unveiled an AI-powered inference appliance for airlines at HUAWEI CONNECT 2026. Huawei says it handles tasks including NOTAM parsing, aircraft maintenance translation and safety report analysis.

@alibaba_cloud on X published this on Oct 12, 2026 (source: x.com).

At Apsara Conference 2026, Jason Wu, Alibaba Cloud's head of Elastic Compute Service, discussed how AI is changing what infrastructure must deliver, from model training and inference to running AI agents at scale.

@alibaba_cloud on X published this on Oct 12, 2026 (source: x.com).

Alibaba Cloud said QwenWork can produce a ready-to-use audio file from a description such as a genre, vocals, lyrics and song structure, from a single track for one video to a full album.

@InfronAI on X published this on Oct 12, 2026 (source: x.com).

Infron said DeepSeek V4.1 Flash is up to 80% off on Infron Batch, priced at $0.06 per million input tokens and $0.24 per million output tokens with a 1M-token context.

@elonmusk on X published this on Oct 12, 2026 (source: x.com).

Elon Musk shared a Sawyer Merritt post saying Grok Bot can now be used in a Tesla Cybercab. Per the post, riders sign into Grok once, their Bots are available on later rides and are logged out when they leave the car, and the agent can add items to an Amazon cart, send emails, order coffee and manage calendars by voice.

@alibaba_cloud on X published this on Oct 12, 2026 (source: x.com).

Alibaba Cloud said users can tell QwenWork once to check competitor prices every Monday at 9 AM. It opens each page, compares against the previous week's snapshot, and reports new lines, markdowns and promotions with timestamped screenshots.

@alibaba_cloud on X published this on Oct 11, 2026 (source: x.com).

Alibaba Cloud said users can drop a reference video into QwenWork to break down its creative elements, including pacing, transition patterns, text timing and overall structure, to guide their own product videos.

@alibaba_cloud on X published this on Oct 11, 2026 (source: x.com).

Alibaba Cloud said users can drop a meeting recording into QwenWork, which turns it into to-dos, using the example of a 52-minute holiday campaign meeting.

Anvisha (Moda) said this on Oct 11, 2026; not an official company announcement (source: x.com).

Anvisha posted that Claude Opus was asked to read recent SIGGRAPH papers and build the coolest thing it could. The post says the result is a differentiable particle simulation run backwards that uses NVIDIA Warp to optimize 450,000 launch trajectories to match a song's beat drop, with every pixel generated by code rather than video models.

@alexandr_wang on X published this on Oct 12, 2026 (source: x.com).

Meta Chief AI Officer Alexandr Wang posted that Meta is working to fix a bug in how Muse Wonder and Muse Wonder Full count against usage limits for Muse Code subscribers. He said the behavior is not intended and that Meta will update users once it is fixed and issue usage resets. The post gives no timeline for the fix or the resets.

@elonmusk on X published this on Oct 12, 2026 (source: x.com).

Replying to Emad Mostaque, who said he hit the usage limit on a heavy Grok Bot account days before the week ended despite offloading harder work to Cursor, Devin, Claude and Codex, Elon Musk asked whether Grok already provides an easy way to increase the usage limit. Mostaque also argued that dashboards showing where quotas go should be standard.

Anthropic published this on Oct 12, 2026 (source: anthropic.com).

Anthropic published research on stronger defenses against illicit distillation of Claude. It says early evidence suggests the new methods cut the amount of data distillers harvest by around 90 to 99 percent, including preserved thinking that blocks editing prior context for newer API accounts, a two-stage classifier on leading Claude models with under 0.1% false positives, and bans plus charging for blocked distillation requests. The post cites Zhipu’s GLM-5.3 and Moonshot’s Kimi as examples of distillation campaigns and notes trade-offs for legitimate integrations.

Stanfordnlp published this on Oct 12, 2026 (source: github.com).

Stanford NLP released DSPy 3.4.1. The release makes GEPA’s structured InstructionProposer the default instruction proposer, lets dspy.RLM build DSPy sub-agents from inside its REPL, adds Python 3.15 support, fixes a 3.4.0 import crash, and includes streaming, adapter and interpreter fixes. Upgrade note: GEPA’s default reflection request shape changed, so optimization results can differ from 3.4.0 even with the same seed.

Langchain-ai published this on Oct 12, 2026 (source: github.com).

LangChain AI released LangGraph 1.2.15. The maintenance release assigns an id to every message written to a messages DeltaChannel, saves a cached node’s writes like a node that ran, stops at an interrupt_before breakpoint when a drain is requested before it, keeps an exit run from a checkpoint the thread moved past out of its newer branch, and stores an update_state input’s writes on the correct checkpoint.

R/LocalLLaMA published this on Oct 12, 2026 (source: reddit.com).

Intel published the GPU D-Series product page for its Crescent Island inference accelerator, headlined by the GPU D270. The air-cooled PCIe card is rated at 350 W with up to 480 GB LPDDR5X, 1.5 TB/s memory bandwidth, 32 Xe3P cores, 256 XMX engines and up to 1.3 PFLOPS of FP8 compute, aimed at agentic AI, long contexts and high-concurrency serving without HBM.

Atlas of Applied AI published this on Oct 12, 2026 (source: gravitypersists.github.io).

A developer published Atlas of Applied AI, a navigable map that places applied AI engineering techniques and approaches near what they grew out of. Claude Opus drafted the regions and lineage, Gemini Deep Research expanded each region, and a final Opus pass merged the results. The project is on GitHub Pages for browsing and contributions.

Posted on arXiv on Oct 12, 2026 (source: arxiv.org).

SenseTime, Tsinghua and NTU researchers propose Looped Diffusion Transformer (Looped-DiT), which reuses shared Transformer blocks several times inside each denoising step. With deep supervision and self-modulating attention, a 260M-parameter looped model surpasses a non-looped model about 6.5× larger across text-to-image benchmarks while using about 4.9× less inference compute; loop depth can also be changed at sampling time. Paper: arXiv 2609.40305.

Not confirmed by the company — reported on Oct 13, 2026 (source: wccftech.com).

YouTuber Jun Yamog mounted four 32GB Tesla V100s on an NVLink carrier board, with A100 heatsinks milled down to fit, and ran Qwen3.8-Flash-Next at 4-bit. On long-prompt processing the 2017-era cards were 40% faster than a pair of DGX Sparks and 230% faster than an M5 Ultra, and in short-prompt token generation they hit 100.6 tok/s versus 44.5 tok/s on the M5 Ultra. Yamog attributes much of the Apple gap to missing software optimizations.

Not confirmed by the company — reported on Oct 12, 2026 (source: arstechnica.com).

Bloomberg reports that Epic's Marvel and Star Wars shooter with Disney, slated for March 2027, is delayed indefinitely and moving to an 'AI first' process, with code written by models such as Anthropic's Claude ('no more hand-written code') and art generated with full AI workflows. CEO Tim Sweeney wrote on X that Epic has a huge pivot underway, moving developers around to retool for maximum AI acceleration, with the aim of building more and better rather than doing the same with fewer people, and bringing the benefits to all Unreal teams.

Not confirmed by the company — reported on Oct 12, 2026 (source: argmin.net).

Berkeley's Ben Recht examined the faster-FFT result in OpenAI's recent math paper release. He argues the hidden constant is around 10^100, so Cooley-Tukey stays faster for any realistic input, and that the proof assumes exact complex arithmetic on numbers that grow superpolynomially, a model no real computer matches. He calls the result an artifact of loose formalization rather than a breakthrough and makes a similar critique of the O(n^2.25) matrix multiplication result in the same release.

GitHub published this on Oct 13, 2026 (source: github.blog).

A new telemetry.capture.identity managed setting adds user.name, process.user.name and host.name to Copilot's OpenTelemetry data, so enterprise administrators can attribute activity to a user without collecting prompts, responses or tool content. Admins can also separately enable five capture classes (prompts, responses, toolArguments, toolOutput, policyDetail) in Copilot CLI, the GitHub Copilot app and VS Code Agent Host sessions. Everything is off by default, and managed values override local settings, SDK configuration and environment variables.

Microsoft Foundry Blog published this on Oct 12, 2026 (source: techcommunity.microsoft.com).

A Microsoft Foundry blog post walks through building hosted agents in Foundry Agent Service, published to Teams and Microsoft 365 Copilot, that retrieve SharePoint content as the signed-in user rather than as the agent. The agent keeps its own identity for model inference, while a Foundry toolbox and a UserEntraToken connection pass the user's Entra ID token to a Foundry IQ knowledge base, which queries SharePoint live through the Microsoft 365 Copilot Retrieval API so answers only cite documents that user can open. Remote SharePoint knowledge sources in Foundry IQ are in preview, and a sample is on GitHub.

Microsoft Foundry Blog published this on Oct 12, 2026 (source: techcommunity.microsoft.com).

A Microsoft Foundry blog post explains the official Skills extension for Model Context Protocol (io.modelcontextprotocol/skills, proposal SEP-2640, now Final), which lets an MCP server list the skills it offers with a manifest of files, sizes and SHA-256 digests so compatible hosts can discover and load them instead of keeping per-agent copies. The extension does not change the skill format or run skills on the server. Microsoft says Foundry's skills support, in preview, follows SEP-2640: skills uploaded to a Foundry project can be referenced in a toolbox and loaded by agents through the toolbox's MCP endpoint.

NVIDIA Blog published this on Oct 13, 2026 (source: blogs.nvidia.com).

An NVIDIA blog post lists French companies using its hardware and open models. Bull is extending European AI server production to Vera Rubin Ultra NVL72 systems planned for 2028, the Altair-1 satellite reached orbit running Mistral models on Jetson AGX Orin modules, and LINAGORA and Docaposte's Nemotron-based Twake.ai became the referenced sovereign digital workplace for more than 2,500 French public bodies. Other examples include L'Oréal adopting NVIDIA ALCHEMI for in-silico formulation research and Mirakl onboarding more than 70 million products with open models on NVIDIA GPUs.

Cline published this on Oct 13, 2026 (source: github.com).

Cline CLI v3.0.71 runs plugins inside the hub instead of a separate node or bun process, and plugins that fail to load now raise a warning and appear in a new `cline hub plugins` status view instead of silently losing their tools. The --thinking flag now works in ACP mode, provider API keys set as environment variables count as configured at startup, and the Cline provider defaults to Claude Sonnet 5.5 when no model is chosen. The release also fixes queued scheduled runs that stayed pending, loads symlinked rule files, discovers skills in .claude/skills, and adds the AI/ML API, LowRouter and Prime Inference providers.

Manus published this on Oct 13, 2026 (source: manus.im).

Manus is integrating its agentic capabilities into OPPO AI Mind Pilot's new Agent Mode, debuting on the upcoming OPPO Find X10 series with ColorOS 17. The agent can break a goal into steps and work with files and photos on the phone, for example assembling a day-by-day itinerary from bookings scattered across apps or building a presentation from on-device files and web information. Manus says local tasks run with on-device models and the content is not shared with Manus, while external processing requires explicit user confirmation. Availability, markets and pricing will be announced at the product launch.

Not confirmed by the company — reported on Oct 13, 2026 (source: qbitai.com).

EBRUN will hold its 2026 Industrial Internet annual conference together with its first Industrial Agents Conference on November 20 at the China World Hotel in Beijing, under the theme of modal evolution driven by AI and going global. Sessions will cover deploying industrial AI agents, AI in supply-chain procurement, data and token economics, and overseas expansion. EBRUN will also release its 2026 industrial internet report and an award case collection at the event. The article was provided by EBRUN and republished by QbitAI.

Not confirmed by the company — reported on Oct 13, 2026 (source: qbitai.com).

Alibaba's Qwen app has rolled out Qwen Go, a personal agent feature, to some users in a gray test. In cases shared by testers, it logged into a ticketing platform, found a delayed flight, contacted human customer service and negotiated a 70-yuan voucher into cash, and claimed a 52-yuan price-protection refund on an e-commerce order. It also books car servicing and sports courts and compares appliance prices across subsidies and coupons. QbitAI says it runs on Qwen 3.8's agentic capabilities and draws on Alibaba services such as Taobao, Fliggy and Amap.

Not confirmed by the company — reported on Oct 13, 2026 (source: qbitai.com).

Chinese AI infrastructure company Infinigence AI (Wuwen Xinqiong) has closed a 2.3 billion yuan Series C, taking its total funding to 4.3 billion yuan, which QbitAI calls a record for the domestic AI-native infrastructure sector. The company sits between models and chips, pooling heterogeneous compute from data centers it does not own; it says its platform supports 16 chip types including NVIDIA, AMD and Huawei Ascend and serves 37,000 petaflops of compute. Investors over its rounds include Tencent, Baidu, Lenovo, Sequoia China and state-backed funds.

Not confirmed by the company — reported on Oct 12, 2026 (source: qbitai.com).

Humanoid robot maker UBTech placed 11th with a score of 115 on the 2026 Fortune Future 120, compiled by Fortune with BCG from more than 3,500 companies, ahead of NVIDIA (16th) and OpenAI (18th). UBTech says its first-half 2026 revenue rose 104.2% to 1.27 billion yuan and humanoid robot sales grew 268.3%. The article was provided by UBTech and republished by QbitAI.

Not confirmed by the company — reported on Oct 12, 2026 (source: qbitai.com).

Citing Financial Times and PitchBook data, QbitAI reports that robotics and physical AI companies raised nearly $48 billion globally this year as of October 8, more than in all of 2025. Large rounds included Anduril's $5 billion, NEURA Robotics' up to $1.4 billion Series C, Skild AI's $1.4 billion and Physical Intelligence's $1 billion. PitchBook counted $18.6 billion in the second quarter alone, up 25.3% from the previous quarter even as deal count fell 14.1%.

Not confirmed by the company — reported on Oct 13, 2026 (source: techcrunch.com).

Amex's new framework captures and authenticates a customer's purchase intent before an AI agent completes a transaction, starting with a Create Intent API that records the instructions and constraints such as merchant, maximum budget and product preferences. After the customer authenticates and signs off, Amex creates an Intent ID and a Proof of Intent Token, which is used to obtain a secure payment credential the agent uses at checkout instead of the real card number. The framework is part of Amex's ACE Developer Toolkit, launched in April, and Amex says it hopes it becomes an industry standard. Amex says it has not yet seen an increase in disputes from agentic commerce.

MagicLab said this on Oct 13, 2026; not an official company announcement (source: qbitai.com).

Chinese robotics company MagicLab says its VLM agent framework EchoPolicy-0 ranks first on the MolmoSpaces Combined leaderboard, the Allen Institute for AI's zero-shot embodied AI benchmark, with a 75.4% average success rate across all four v1 manipulation tasks over 3,915 episodes. Its 3.4B-parameter MagicBrain framework ranks fourth at 72.4% and first on the Pick and Pick & Place tasks, with a 94.4% pick success rate. EchoPolicy-0 has a VLM choose among several candidate trajectories proposed by an action model and verifies the result after execution, while MagicBrain marks targets visually in the VLA's observations. The rankings are the company's figures as of October 13.

Not confirmed by the company — reported on Oct 12, 2026 (source: bloomberg.com).

DeepSeek plans to lead a seed round of tens of millions of dollars in Winverse AI, a Chinese startup founded this year to build AI for robots, valuing it at about $400 million after the investment, according to Bloomberg. Founder Hu Han, a co-author of the Swin Transformer paper, previously led Tencent's vision-language and embodied foundation model teams. Meituan's Long-Z Investments has also committed, and the deal is not final. QbitAI notes it would be DeepSeek's second robotics investment in two months, after it took about 141 million yuan of shares in Unitree's IPO placement in August.

Manning published this on Oct 13, 2026 (source: manning.com).

Manning has released AI Recommender Systems: Transformers, LLMs, agents, and RAG by Kim Falk in its MEAP early-access program, with 4 of 14 chapters available and print publication estimated for spring 2027. The book covers building an AI-first recommender system, candidate, filtering, scoring and reranking pipelines, cold-start problems, and deploying and monitoring recommenders in production. Falk, a principal recommender engineer, previously wrote Practical Recommender Systems.

Amazon (Kiro) published this on Oct 12, 2026 (source: kiro.dev).

Amazon's Kiro is now in scope for FedRAMP Class C in the US East (N. Virginia) Region. US government agencies, contractors and other organizations whose workloads must meet FedRAMP Class C requirements can now use Kiro for AI-assisted software development. AWS lists current Kiro scope details on its services-in-scope by compliance program page.